Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

No Experts, No Problem: Avoidance Learning from Bad Demonstrations

Authors: Huy Hoang, Tien Mai, Pradeep Varakantham

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The efficacy of our method is evaluated across standard benchmark environments, where it consistently outperforms state-of-the-art baselines.
Researcher Affiliation Academia Huy Hoang Singapore Management University Institute for Infocomm Research, A*STAR, Singapore EMAIL Tien Mai Singapore Management University EMAIL Pradeep Varakantham Singapore Management University EMAIL
Pseudocode Yes A Pseudo Code of UNIQ Below we present a pseudo code of our UNIQ algorithm. Algorithm 1 UNIQ: UNdesired Demonstrations driven Inverse Q-Learning
Open Source Code Yes We have describe how to generate our data as well as provide it along with our submitted source code with sufficient instructions for their use.
Open Datasets Yes We evaluate our method on two popular benchmark environments, Mujoco and Safety-Gym, using public datasets [38, 21]. Our experiments demonstrate superior performance compared to several state-of-the-art baselines. ... For the Mujoco experiments, we use the official D4RL dataset [10], which consists of three performance levels: random, medium, and expert.
Dataset Splits Yes For the Mujoco tasks, we use 5 trajectories each from the random and medium datasets to construct the undesirable dataset DUN, and 500, 500, and 100 trajectories from the random, medium, and expert datasets, respectively, to construct the unlabeled dataset DMIX.
Hardware Specification Yes Our experiments were carried out on a GPU cluster equipped with 8 NVIDIA RTX 3090 GPUs. Each experimental setup was run with five distinct training seeds in parallel, sharing a single GPU, eight CPU cores, and 64 GB of RAM.
Software Dependencies Yes The software environment was built using JAX version 0.4.28, with support for CUDA 12 specifically, CUDA version 12.3.2 and cu DNN version 8.9.7.29.
Experiment Setup Yes For fair comparison, we keep the same basic hyper-parameters across all the baselines which are detailed as follow for Mujoco, Safety-gym and Mujoco-velocity tasks as Table 4: